Mapping and predicting cassava mosaic disease outbreaks using earth observation and meteorological data-driven approaches

Abstract

Cassava Mosaic Disease (CMD) poses a significant threat to cassava (Manihot esculenta Crantz) production, both globally and in Thailand. Caused by the Cassava Mosaic Begomovirus (CMB), CMD conventional detection methods involve field-based visual assessment and mapping, which can be time-consuming and labour-intensive.In this study, field-based surveys have been conducted during the early wet season cassava planting period in three different productive regions of Thailand. Five classification models (namely ANN, RF, XGBoost, SVM, and Ensemble learning) were employed to predict CMD occurrence with various sets of optimized predictive meteorological and remote sensing features. Field based CMD observations have been used for modelling validation and accuracy assessment.Utilizing time series of Sentinel-2 MSI Level-2A images, we estimate cassava start of season (SOS) across both before and after planting periods on a gridded basis. Establishing meteorological and remote sensing time frames, we aggregate predictive features through different temporal windows. Employing interactive and multistage correlation analysis, along with feature importance ranking, we optimize the selection of predictive features, effectively reducing multicollinearity. The Ensemble model and the usage of combined meteorological and remote sensing optimized predictive features showed best prediction results achieving an overall accuracy of 90%, F1-score of 0.90, Matthews Correlation Coefficient of 0.71, unit-normalized MCC* of 0.85, and an average central actual accuracy of approximately 0.85.Observed CMD patterns demonstrate that soil-crop suitability and farming practices (including the use of most CMD susceptible cultivars) significantly influence the CMD spread in selected farming areas. However, modelling analysis revealed that, among several meteorological features, CMD occurrence is associated with predominant winds blowing across the entire before and after planting periods, mainly from the first quadrant, and to a lesser extent, humidity. Notably, higher disease incidence areas corresponded with SOS events in January and February. In predicting CMD, individual S2-MSI bands were not crucial, but NDVI and EVI emerged as the most explanatory remote sensing features, followed by GNDVI and PSRI indices.The study's findings highlight that coupling CMD occurrence modelling, combined remote sensing and meteorological data with conventional field-based observations may significantly enhance targeted disease management protocols and assessing potential impacts on cassava yields.